Adding AI to an edge device usually forces a choice. You can replace the processor you already ship with an AI-enabled MCU or SoC, which means a board respin, new firmware and a new qualification cycle. Or you can add an accelerator built for cameras and robots, and budget watts of power and a heatsink for it, which can mean redesigning the power supply, thermal design and form factor.

AKD1500 gives you another option: add a dedicated AI co-processor next to the host you already have. Teams building continuous sensing or streaming products keep the host architecture that serves the product, and add AI processing where it is needed.

Available in PCIe and SPI variants, AKD1500 brings BrainChip’s event-based Akida architecture and on-device learning to designs ranging from MCU-based sensing products to application-processor systems.

As a co-processor, AKD1500 splits the work: your host keeps running the application, and AKD1500 runs the neural network and returns the result. In a smart speaker, for example, the MCU keeps handling audio capture, connectivity and the app, and passes microphone data to AKD1500 for keyword spotting.

It comes in two configurations to match your host:

• AKD1500 PCIe pairs with the application processor in systems like industrial PCs, edge servers and robots. It’s built for performance and scale, and can use host memory for larger models, which suits vision, multi-modal models and always-on anomaly detection.
• AKD1500 SPI pairs with the low-power MCU already inside battery-powered and embedded devices, adding AI without an MCU upgrade. It suits wearables monitoring vital signs, smart home devices listening for keywords, and industrial IoT nodes detecting vibration anomalies at the sensor.

If you’d rather start on a board, AKD1500 is also available on the PCIe Development Card and the M.2 Card. Explore more

Whichever configuration you choose, the chip brings the same two strengths: low-power inference driven by activation sparsity, and on-chip learning that adapts to new users and conditions after your product ships.

Low-Power AI Driven by Sparsity

Edge sensing products are measured differently from benchmark demos. A keyword-spotting device has to stay ready for speech, and a machine monitor has to recognize the sensor pattern that signals a fault. What matters is meeting your accuracy and response-time targets within the device’s power budget, all day.

AKD1500 does this with an event-based neuromorphic architecture. It takes advantage of activation sparsity, the inactive values inside a neural network, and skips them instead of computing them, which cuts processing activity and power.

On-Device Learning for Personalized Edge AI

Some products need to do more than recognize the classes defined before shipment. They need to adapt to a user or deployment.

AKD1500 includes on-device learning capabilities for application personalization. A concrete example is AkidaTag’s keyword-learning demonstration: users can teach the device a new keyword through its companion app, with learning performed on-chip rather than through a cloud retraining pipeline.

This gives developers a practical way to explore personalized voice interfaces without making cloud connectivity a requirement for that learning step.

AkidaTag makes the host-plus-co-processor approach tangible.

The battery-powered reference platform combines an AKD1500 with Nordic Semiconductor’s nRF5340 wireless SoC. Its firmware drives AKD1500 over SPI, while the platform provides sensor handling, Bluetooth connectivity, and a companion application. Explore documents and pre-order now

For engineers exploring the SPI path more directly, BrainBoard1500 offers another starting point. Designed and manufactured by Neuromorphyx around AKD1500, the developer and evaluation board provides a direct SPI/QSPI path for neural-network processing on sensor data. Explore more

These platforms illustrate the same design principle: a host manages the device, while a dedicated co-processor handles AI.

Start with a supported model and a clear application requirement.

AKD1500 is a strong fit if your product:

• has to sense continuously, like listening for a wake word or monitoring a machine for vibration
• runs on a battery or a fixed power budget
• already has an MCU or application processor you want to keep
• needs to adapt to a user or site after it ships

Voice control, predictive maintenance, health monitoring, and edge vision are among the family’s target applications.

The best way to know is to test your own model. Start from a ready-made model on GitHub, such as keyword spotting, or convert yours with MetaTF. Then check accuracy and latency on the AKD1500 platform that matches your host.

The AKD1500 family provides several ways to begin:

• The AKD1500 Raspberry Pi 5 Dev Kit for an evaluation starting point.
• M.2 and PCIe development cards for compatible host systems.
• BrainBoard1500 lets you prototype over SPI on a partner board
• AkidaTag for a compact, connected sensing reference platform.
• AKD1500 co-processor silicon for custom hardware integration.

BrainChip’s MetaTF development flow supports compiling and optimizing models for AKD1500, while the PCIe and SPI variants provide different host-integration paths.

Stay tuned for an AKD1500 family buying guide to choose the right interface and development platform, or explore the family in the BrainChip Shop https://shop.brainchipinc.com/.